Integrating AI Agents into Companies
AI needs a different environment than humans to take advantage of its speed.
The AI Management Challenge
Recent AI models have significantly improved their capability, creating enormous potential for businesses and organizations that can harness this capability. But these models aren't straightforward to integrate into "meat space." At this moment, there are two main differences between AI models and humans:
AI models are fast.
A project that might take a competent human a day can be done by AI models in a few minutes, as long as they have the context. Their input/output speed and recall are orders of magnitude better than what humans can do.
Models lack some human context.
Social cues, organizational hierarchies, and the physical world can challenge models. The models need this information to be more effective. We should expect significant improvement in these domains, but it may take time.
The Tenants of AI Management Philosophy
Organizations must use AI's speed and provide context efficiently to unlock productivity gains. There also needs to be a framework that can maintain quality even at higher speeds. Several strategies jump out:
Massively increase the use of wikis and other written content.
Human organizations rarely codify their entire structure because the upfront cost and coordination are substantial. The ongoing effort to access and maintain such documentation is also significant. Asking co-workers questions or developing working relationships is usually more efficient and flexible.
Asking humans or developing relationships nullifies AI's strength (speed) and exposes its greatest weakness (human context). Having the information in written form eliminates these issues. The cost of creating and maintaining these resources should fall with the help of AI.
I've written about how organizations already codify themselves as they automate with traditional software. Creating wikis and other written resources is essentially programming in natural language, which is more accessible and compact.
Move from reviews to standardized pre-approvals and surveillance.
Human organizations often prefer reviews as a checkpoint because creating a list of requirements is time-consuming, and they are commonly wrong. A simple review and release catches obvious problems and limits overhead and upfront investment. Reviews of this style are still relevant for many AI tasks where a human prompts the agent and then reviews the output.
AI could increase velocity for more complex and cross-functional projects by moving away from reviews. Waiting for human review from various teams is slow. Alternatively, AI agents can generate a list of requirements and unit tests for their specialty in a few minutes, considering more organizational context (now written) than humans can. Work that meets the pre-approval standards can continue, and then surveillance paired with graduated rollouts can detect if there are an unusual amount of errors.
Human organizations have a tradeoff between "waterfall" and "agile," AI organizations can do both at once with minimal penalty, increasing iteration speed.
Use "Stop Work Authority" methods to ensure quality.
One of the most important components of the Toyota Production System is that every employee has "stop work authority."" Any employee can, and is encouraged to, stop the line if they see an error or confusion. New processes might have many stops as employees work out the kinks, but things quickly line out. It is a very efficient bug-hunting method.
AI agents should have stop work authority. They can be effective in catching errors because they work in probabilities. Work stops when they cross a threshold of uncertainty. Waymo already does this with AI-driven taxis. The cars stop and consult human operators when confused.
An obvious need is a human operations team that can respond to these stoppages in seconds or minutes.
Issues are recorded and can be fixed permanently by adding to written context resources, retraining, altering procedures, or cleaning inputs.
Design for AI.
A concept called "Design for Manufacturing" is popular with manufacturing nerds and many leading companies. The idea is that some actions are much cheaper and defect-free than others. For instance, an injection molded plastic part with a shape that only allows installation one way will be a fraction of the cost of a CNC-cut metal part with an ambiguous installation orientation. The smart thing to do is design a product to use the plastic part instead of a metal one.
The same will be true of AI agents. Designing processes for their strengths will have immense value, especially in production, where errors are costly.
Cast a Wider Design Net.
The concept of "Design for AI" also applies at higher levels. Employees with the creativity for clever architectural designs are scarce resources. AI agents can help by providing analysis of many rabbit holes and iterations, helping less creative employees or supercharging the best.
The design phase has the most impact on downstream cost and productivity of any phase.
Source: Munro and Associates
Minimize human touch points.
Human interaction significantly slows down any process and kills one of the primary AI advantages.
Written context is the first step in eliminating human touch points. Human workers can supervise the creation of the wikis instead of completing low-level work.
Pre-approvals are the next, so AI agents are not waiting for human sign-off.
AI decision probability thresholds, graduated rollouts, and unit tests can reduce the need for human inspection of work output.
Eliminate meeting culture.
Meetings help human organizations coordinate tasks and exchange context. Humans will continue to have meetings even in AI organizations.
The vast majority of lower-level meetings need to be cut. They lose their advantages once work completion times are compressed and context more widely available.
Meeting content moves from day-to-day operations to much higher-level questions about strategy and coordination. Humans might spend even more time in meetings if the organizational cadence increases so that strategies have to constantly adjust!
Robots?
General-purpose robots should be much more common in the coming years. A robot arm won't be able to work 100x faster than a single human at a specific task. However, supervisors can organize them to allow scaling of resources and compression of total work time by reducing upfront training, coordination, and context-switching costs. For example, you might be able to build a house in one day.
Robots are also important for expanding AI beyond paperwork. Much better (and easier to create) physical models are needed to improve the world of atoms faster. You can ask an AI agent for suggestions on assembling a house faster, but it's likely to lack the detail needed to make a difference. Data from the robots and human assembly workers can improve the fidelity of the digital model. Better models will make the design work much more productive by handling more iteration cycles in simulation.
The AI Organization
The rise of AI organizations looks similar to software-oriented startups but supercharged. Putting the entire context of an organization in text is easier from the start than patching an existing organization. The best companies will treat AI agent waiting time like Toyota does inventory. And having AI speed from the beginning will make iteration cycles much faster.
One of the major themes is that many opportunities have too high fixed or coordination costs for human organizations to take advantage of. The speed and low cost of AI agents make these opportunities viable and allow more extensive design-phase exploration that enables much simpler and more productive daily operations.
Roadblocks outside the organization, like permits, become even more costly to prosperity. But, efforts to bypass these blockages will be much more robust with AI support. The most free and open economies should benefit the most.
We are at the dawn of an exciting era that will likely span decades.